Multiscale exploration of informative latent features for accurate deep eutectic solvents viscosity prediction

Abstract

Deep eutectic solvents (DESs) are promising green solvents, yet their high and variable viscosity presents challenges in practical applications. Traditional viscosity measurements are labor-intensive and time-consuming due to numerous influencing factors. This study introduces a novel prediction framework integrating message passing neural networks (MPNN)-graph attention networks (GAT)-multilayer perceptron (MLP). Using a dataset of 5790 DESs, recognizing the essential role of SMILES in predicting DESs viscosity, two stacked GAT layers were utilized to implicitly capture interdependencies among molecular substructures, enabling the extraction of significant features. Given that DESs are typically binary systems, the predicted density is incorporated as an additional input, reducing reliance on experimental data. The MLP combines these extracted features with physical and chemical properties for accurate viscosity prediction. This multiscale, data-driven approach significantly improves prediction performance (R² = 0.9945, AARD = 2.69%), surpassing conventional methods and advancing green solvent design.